introduced a cache for `followAllReferences()` calls #7192

firewave · 2025-01-07T20:24:06Z

No description provided.

firewave · 2025-01-07T20:28:21Z

This essentially eliminates any meaningful impact by followAllReferences() at all.

-D__GNUC__ --check-level=exhaustive ../lib/utils.cpp

Clang 19 652,987,030 -> ~~624,476,510~~ 618,089,977

followAllReferences() calls from isAliasOf() - 350,100 -> 1,581

The example from https://trac.cppcheck.net/ticket/10765#comment:4:

Clang 19 3,056,382,003 -> ~~2,838,708,731~~ 2,815,165,117

followAllReferences() calls from isAliasOf() - 2,592,565 -> 641

firewave · 2025-01-07T20:33:25Z

lib/valueflow.cpp

@@ -3738,7 +3738,7 @@ static void valueFlowForwardConst(Token* start,
        } else {
            [&] {
                // Follow references
-                auto refs = followAllReferences(tok);
+                auto refs = tok->refs();


needs to be const auto&.

Done.

I filed https://trac.cppcheck.net/ticket/13533 about detecting this.

firewave · 2025-01-07T20:34:32Z

lib/vf_analyzers.cpp

@@ -627,7 +627,7 @@ struct ValueFlowAnalyzer : Analyzer {
        if (invalid())
            return Action::Invalid;
        // Follow references
-        auto refs = followAllReferences(tok);
+        auto refs = tok->refs();


This copy is necessary since an additional entry is being added. But I think this is not necessary and I will try to refactor the code to avoid it.

I adjusted this by unfortunately there is some redundant code introduced.

pfultz2 · 2025-01-07T20:40:42Z

lib/token.cpp

+    if (!mImpl->mRefs)
+        mImpl->mRefs = new SmallVector<ReferenceToken>(followAllReferences(this));
+    return *mImpl->mRefs;
+}


This is not the right way to do this. This is a const method that is modifying the token. Instead followAllReferences should be moved to the SymbolDatabase and there should be a pass that fills this in for all of the tokens.

Yes. I was about to add a comment about this. This violates the technical const and if we would not allow this (I hope some day I will finish up that change) this would require mutable (which from my experience is acceptable for caches inside objects).

I am not sure how easy it would be to implement an earlier pass since it is not done for all tokens but there are lots of checks which are performed before we actually end up following references. That would need to be replicated I reckon - and that also has a certain visible overhead and we would need to run through that twice then.

Actually I would also have the ValueFlow behave this way so we might avoid running it for code which is not relevant.

This is not the right way to do this. This is a const method that is modifying the token.

That should be totally fine (by precedent). We modify const Token objects all over the place in the ValueFlow and symbol database via const_cast. Obviously it would be better if we didn't but here it is much cleaner and in a single place and as stated before I think this is acceptable practice.

Actually I would also have the ValueFlow behave this way so we might avoid running it for code which is not relevant.

Please disregard this. This is wishful thinking as this would not be possible the way the ValueFlow is working. I totally forgot I already looked into this.

The const_cast should be fixed, but we shouldn't add more code that needs to be fixed.

Also this is called in ValueFlowForward and ValueFlowReverse so its already called on almost every token in functionScopes, so it really won't help performance being a cache.

Furthermore, in copcheck we update the tokens through passes rather than using a cache, this makes it easier to debug and we can provide this information to addons later on. So doing a pass in SymbolDatabase would be consistent with the rest of the code.

Furthermore, in copcheck we update the tokens through passes rather than using a cache, this makes it easier to debug and we can provide this information to addons later on. So doing a pass in SymbolDatabase would be consistent with the rest of the code.

Will give it a try and check how it impacts performance.

Do we know if it causes a perf impact or how much? It seems we are making it worse for premature optimizations.

There are other advantages to doing it the correct way too such as better debugging and addons can take advantage of this information (this seems like a useful analysis for addons). So if we enable it for addons then we will beed to run a pass regardless.

Also you could consider skipping this for functions we are skipping analysis for, if the performance is too bad, but it would be good to see some actual numbers to make this decision.

So the current approach seems like the best approach.

I meant to say "It seems like the currently best approach".

Do we know if it causes a perf impact or how much? It seems we are making it worse for premature optimizations.

Various performance numbers are in the PR. It is a massive improvement. It would also help with the runtime of the CI.

Also you could consider skipping this for functions we are skipping analysis for, if the performance is too bad, but it would be good to see some actual numbers to make this decision.

That was an idea regarding the ValueFlow (see https://trac.cppcheck.net/ticket/12528) but that won't work since not all passes are based on function scopes. But that is currently out-of-scope and is something I am looking at within another context hopefully soon.

It might actually not an issue after all because with the duplicated calls eliminated it basically no longer has any footprint. The only issue could be that we perform it for more tokens than we actually need so that would introduce new overhead but it might not be much. Will test that. Although I would prefer not to have that at all since all the overhead adds up - a lot.

I just realized this is called when setting exprids, so it always called on every token regardless of ValueFlow analysis.

I just gave it a small spin and it does not increase the Ir so that looks feasible.

pfultz2 · 2025-01-07T20:43:40Z

lib/token.h

@@ -143,6 +145,8 @@ struct TokenImpl {
    void setCppcheckAttribute(CppcheckAttributes::Type type, MathLib::bigint value);
    bool getCppcheckAttribute(CppcheckAttributes::Type type, MathLib::bigint &value) const;

+    SmallVector<ReferenceToken>* mRefs{};


If this is going to be a pointer, you should use std::unique_ptr or std::shared_ptr.

I modeled it after mValues which is also just a raw pointer.

You should model it after mAttributeAlignas, thats a much safer choice. I think the raw pointers were used before managed pointers were available.

That requires ReferenceToken to be a complete type and the astutils.h to be added.

firewave · 2025-01-07T20:59:27Z

Something really weird is going on here in the UBSAN job:

Check time: cli/threadexecutor.cpp: 0.53017s
Check time: cli/processexecutor.cpp: 1.41327s
Check time: lib/addoninfo.cpp: 0.172107s
Check time: lib/analyzerinfo.cpp: 0.636273s

The timing information for cli/cmdlineparser.cpp is missing ...

firewave · 2025-01-07T21:20:25Z

Before

Check time: cli/cmdlineparser.cpp: 1091.73s
[...]
Check time: lib/checkio.cpp: 219.069s
[...]
Check time: lib/symboldatabase.cpp: 191.785s
[...]
Check time: lib/tokenize.cpp: 290.026s

After

Check time: cli/cmdlineparser.cpp: 760.299s
[...]
Check time: lib/checkio.cpp: 168.103s
[...]
Check time: lib/symboldatabase.cpp: 145.913s
[...]
Check time: lib/tokenize.cpp: 236.561s

firewave · 2025-03-15T09:15:58Z

Could we merge this without the debug integration for now (I will file a ticket about doing that)? I would like to clean up and test the existing debug output first (see #7258 - as usual stalled) before I add new data. And it would greatly speed up the CI as well as giving a baseline to compare against if running it as a pass would have significant performance impact.

pfultz2 · 2025-04-05T17:04:55Z

This should be done in the symboldatabase either before or during exprids.

pfultz2 · 2025-04-05T17:07:01Z

lib/vf_analyzers.cpp

+                return a;
+        }
+        if (analyzeTok) {
+            auto a = analyze_f(tok);


I dont really like this at all, it is harder to follow and harder to extend in the future.

I dropped it for now.

pfultz2 · 2025-04-05T17:18:36Z

lib/astutils.h

-                                                bool inconclusive = true,
-                                                ErrorPath errors = ErrorPath{},
-                                                int depth = 20);
+SmallVector<ReferenceToken> followAllReferences(const Token* tok, bool temporary = true);


Since we are removing the inconclusive flag, then we should add that to the ReferenceToken so we can know if it was inconclusive or not.

The removed flags were not used outside of astutils.cpp itself. If that information is need in the future we can make the suggested change.

pfultz2 · 2025-04-05T17:19:21Z

lib/token.cpp

@@ -2717,3 +2718,10 @@ const Token* findLambdaEndScope(const Token* tok) {
 const std::string& Token::fileName() const {
    return mTokensFrontBack.list.getFiles()[mImpl->mFileIndex];
 }
+
+const SmallVector<ReferenceToken>& Token::refs() const


This is missing the bool temporary parameter.

That was intentional - added.

firewave · 2025-04-07T09:06:22Z

Thanks for the feedback. Some of the shortcomings were intentional since it seemed to make things simpler and I wanted to get the performance improvement in. But I remembered the changes to be simpler which is not the case so I need a bit to get into it again.

firewave · 2025-05-07T06:42:23Z

@pfultz2 How should the information be presented in the debug output? I also assume it should be a separate section.

firewave · 2025-06-03T12:26:01Z

This should be done in the symboldatabase either before or during exprids.

In the limited cases I tested that produced considerably more calls to followAllReferences() but did not affect the overall performance much. This still needs to be tested with the selfcheck.

…for the GUI (#7570)

firewave · 2025-08-14T09:18:09Z

Given the improvement this change provides and how it affects the CI (we are actually getting even slower and not having this applied makes it harder to pinpoint the other hot spots) I would really prefer if we could just merge the very first version and do the adjustments as a follow-ups.

firewave · 2025-08-14T09:27:21Z

Having this done in incremental commits would also make it possible to bisect differences in behavior/performance this might have based on the stage this cache is generated. Especially after I tried to add a different cache in #7573 which relies on that the results are not being cached beyond a point (which might also cause issues here if this is stored too early - but might not be reflected in our tests).

firewave commented Jan 7, 2025

View reviewed changes

pfultz2 reviewed Jan 7, 2025

View reviewed changes

firewave force-pushed the followref-cache branch from 42d1ec9 to b38e8ba Compare January 8, 2025 13:04

firewave marked this pull request as ready for review January 8, 2025 13:38

firewave force-pushed the followref-cache branch from b38e8ba to a36380e Compare January 14, 2025 15:38

firewave marked this pull request as draft January 19, 2025 16:48

firewave force-pushed the followref-cache branch from a36380e to f54680c Compare April 5, 2025 08:08

firewave marked this pull request as ready for review April 5, 2025 08:08

pfultz2 reviewed Apr 5, 2025

View reviewed changes

firewave marked this pull request as draft April 7, 2025 07:33

firewave force-pushed the followref-cache branch 2 times, most recently from fb74dde to 6699af7 Compare June 3, 2025 12:18

firewave force-pushed the followref-cache branch from 6699af7 to 4e3a9ce Compare June 4, 2025 10:09

firewave changed the title ~~introduced a cache for followAllReferences() calls with default parameters~~ introduced a cache for followAllReferences() calls Jun 4, 2025

firewave added 3 commits June 5, 2025 23:50

introduced a cache for followAllReferences() calls [skip ci]

f233e85

generate token references during symbol database creation [skip ci]

7627961

s [skip ci]

270a9c8

firewave force-pushed the followref-cache branch from 4e3a9ce to 270a9c8 Compare June 5, 2025 22:01

firewave referenced this pull request Jul 14, 2025

fixed #13909 / refs #13914 - some (enforced) language handling fixes …

1af769e

…for the GUI (#7570)

introduced a cache for followAllReferences() calls #7192

Are you sure you want to change the base?

introduced a cache for followAllReferences() calls #7192

Uh oh!

Conversation

firewave commented Jan 7, 2025

Uh oh!

firewave commented Jan 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment was marked as duplicate.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

firewave commented Jan 7, 2025

Uh oh!

firewave commented Jan 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

firewave commented Mar 15, 2025

Uh oh!

pfultz2 commented Apr 5, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

firewave commented Apr 7, 2025

Uh oh!

firewave commented May 7, 2025

Uh oh!

firewave commented Jun 3, 2025

Uh oh!

firewave commented Aug 14, 2025

Uh oh!

firewave commented Aug 14, 2025

Uh oh!

Uh oh!

introduced a cache for `followAllReferences()` calls #7192

introduced a cache for `followAllReferences()` calls #7192

firewave commented Jan 7, 2025 •

edited

Loading

firewave commented Jan 7, 2025 •

edited

Loading